[svn-r11231] Purpose:

bug fix for collective chunk IO, phase 1
Optimization hasn't been done yet, the collective chunk IO bug should be fixed.
Description:
In chunking storage, memory space and file space will be remapped, So to check
whether file space and memory space are regular in order to use optimized MPI derived
datatype for collective call one has to check per-chunk wise instead of per hyperslab wise.
Even a regular memory space will be stored in span-tree and will be irregular before chunk IO.

Solution:
1. Check file space and memory space per chunk wise instead of per hyperslab wise.
2. For collective IO mode, number of chunks covered by hyperslab may be different. Since we are
handing per chunk per IO, for the extra chunk IO for some(not all) processors, collective mode will
cause program hanged. So for the extra chunk Io mode independent IO has to be used.
3. On some platforms, Complex MPI derived datatype is not working, so we have to use independent IO for collective IO mode if the selection is irregular. However, when the selection is regular, we do want to use collective IO since that will improve performance. Special cares have to be added for this case.


Platforms tested:
copper(AIX 5.1) Linux(heping mpich 1.2.6), Teragrid machine, Cobalt(altix), modi4

Misc. update:

1 files changed, 40 insertions, 359 deletions

diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c
index e216d85..f85c551 100644
--- a/src/H5Dmpio.c
+++ b/src/H5Dmpio.c
@@ -43,12 +43,6 @@ H5D_mpio_spaces_xfer(H5D_io_info_t *io_info, size_t elmt_size,
                      void *buf/*out*/, 
 		     hbool_t do_write);
 
-/* For irregular hyperslab selection. */
-static herr_t
-H5D_mpio_spaces_span_xfer(H5D_io_info_t *io_info, size_t elmt_size,
-                     const H5S_t *file_space, const H5S_t *mem_space,
-                     void *buf/*out*/, 
-		     hbool_t do_write);
 
 /*-------------------------------------------------------------------------
  * Function:	H5D_mpio_opt_possible
@@ -66,10 +60,10 @@ H5D_mpio_spaces_span_xfer(H5D_io_info_t *io_info, size_t elmt_size,
  *
  *-------------------------------------------------------------------------
  */
+
 htri_t
 H5D_mpio_opt_possible( const H5D_t *dset, const H5S_t *mem_space, const H5S_t *file_space, const unsigned flags)
 {
-    htri_t c1,c2;               /* Flags whether a selection is optimizable */
     htri_t ret_value=TRUE;
 
     FUNC_ENTER_NOAPI(H5D_mpio_opt_possible, FAIL);
@@ -88,16 +82,7 @@ H5D_mpio_opt_possible( const H5D_t *dset, const H5S_t *mem_space, const H5S_t *f
             && (H5S_SIMPLE==H5S_GET_EXTENT_TYPE(file_space) || H5S_SCALAR==H5S_GET_EXTENT_TYPE(file_space))))
         HGOTO_DONE(FALSE);
 
-    /* Check whether both selections are "regular" */
-    /*#ifndef KYANG*/
-#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS
-    c1=H5S_SELECT_IS_REGULAR(file_space);
-    c2=H5S_SELECT_IS_REGULAR(mem_space);
-    if(c1==FAIL || c2==FAIL)
-        HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, FAIL, "invalid check for single selection blocks");
-    if(c1==FALSE || c2==FALSE)
-        HGOTO_DONE(FALSE);
-#endif
+
     /* Can't currently handle point selections */
     if (H5S_SEL_POINTS==H5S_GET_SELECT_TYPE(mem_space) || H5S_SEL_POINTS==H5S_GET_SELECT_TYPE(file_space))
         HGOTO_DONE(FALSE);
@@ -107,126 +92,35 @@ H5D_mpio_opt_possible( const H5D_t *dset, const H5S_t *mem_space, const H5S_t *f
             (flags&H5S_CONV_STORAGE_MASK)!=H5S_CONV_STORAGE_CHUNKED)
         HGOTO_DONE(FALSE);
 
-    if ((flags&H5S_CONV_STORAGE_MASK)==H5S_CONV_STORAGE_CHUNKED) {
-        hsize_t chunk_dim[H5O_LAYOUT_NDIMS];        /* Chunk dimensions */
-        hsize_t startf[H5S_MAX_RANK],      /* Selection start bounds */
-            endf[H5S_MAX_RANK];     /* Selection end bounds */
-        unsigned dim_rankf;         /* Number of dimensions of file dataspace */
-        int pcheck_hyper,check_hyper,   /* Flags for checking if selection is in one chunk */
-            tnum_chunkf,            /* Number of chunks selection overlaps */
-            max_chunkf,             /* Maximum number of chunks selection overlaps */
-            min_chunkf,             /* Minimum number of chunks selection overlaps */
-            num_chunks_same;        /* Flag indicating whether all processes have the same # of chunks to operate on */
-        unsigned dim_chunks;        /* Temporary number of chunks in a dimension */
-        MPI_Comm comm;              /* MPI communicator for file */
-        int mpi_rank;               /* Rank in MPI communicator */
-        int mpi_code;               /* MPI return code */
-        unsigned u;                 /* Local index variable */
-
-        /* Disallow collective I/O if there are any I/O filters on chunks */
-        if(dset->shared->dcpl_cache.pline.nused>0)
-            HGOTO_DONE(FALSE)
-
-        /* Getting MPI communicator and rank */
-        if((comm = H5F_mpi_get_comm(dset->ent.file))==MPI_COMM_NULL)
-            HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve MPI communicator")
-        if((mpi_rank = H5F_mpi_get_rank(dset->ent.file))<0)
-            HGOTO_ERROR(H5E_DATASPACE, H5E_CANTGET, FAIL, "can't retrieve MPI rank")
-
-      /* Currently collective chunking storage 
-	 inside HDF5 is supported for either one of the following two cases:
-	 1. All the hyperslabs for one process is inside one chunk.
-	 2. For single hyperslab selection, the number of chunks that covered 
-	    the single selection for all processes should be equal. 
-	    KY, 2004/7/14
-      */
-
-      /* Quincey, please read.
-	 This is maybe redundant, I think only when both memory and file space be SCALAR
-	 space, the collective IO can work. Otherwise, SELECT_POINT will be reached,collective
-	 IO shouldn't work.
-	 Please clarify and correct the code on the following,
-         Quincey said that it was probably okay if only one data space is SCALAR, 
-         Still keep the code here until we added more tests later.
-	 Kent */
-        if(H5S_SCALAR==H5S_GET_EXTENT_TYPE(mem_space) || H5S_SCALAR ==H5S_GET_EXTENT_TYPE(file_space)) {
-            if(!(H5S_SCALAR==H5S_GET_EXTENT_TYPE(mem_space) && H5S_SCALAR ==H5S_GET_EXTENT_TYPE(file_space)))
-                HGOTO_DONE(FALSE)
-            else
-                HGOTO_DONE(TRUE)
-        } /* end if */
-
-        dim_rankf = H5S_GET_EXTENT_NDIMS(file_space);
-
-        if(H5S_SELECT_BOUNDS(file_space,startf,endf)==FAIL)
-            HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE,FAIL, "invalid check for single selection blocks");
-
-        for(u=0; u < dset->shared->layout.u.chunk.ndims; u++) 
-            chunk_dim[u] = dset->shared->layout.u.chunk.dim[u];
-
-        /* Case 1: check whether all hyperslab in this process is inside one chunk.
-           Note: we don't handle when starting point is less than zero since that may cover
-           two chunks. */
-
-        /*for file space checking*/
-        pcheck_hyper = 1;
-        for (u=0; u<dim_rankf; u++)
-            if(endf[u]/chunk_dim[u]!=startf[u]/chunk_dim[u]) {
-                pcheck_hyper = 0;
-                break;
-            }
-      
-        if (MPI_SUCCESS != (mpi_code= MPI_Reduce(&pcheck_hyper,&check_hyper,1,MPI_INT,MPI_LAND,0,comm)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_Reduce failed", mpi_code)
-        if (MPI_SUCCESS != (mpi_code= MPI_Bcast(&check_hyper,1,MPI_INT,0,comm)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
-
-        /*if check_hyper is true, condition for collective IO case is fulfilled, no
-         need to do further test. */
-        if(check_hyper)
-            HGOTO_DONE(TRUE); 
+    /*The handling of memory space is different for chunking 
+	  and contiguous storage,
+	  For contigous storage, mem_space and file_space won't
+	  change when it it is doing disk IO.
+	  For chunking storage, mem_space will change for different
+	  chunks. So for chunking storage, whether we can use 
+	  collective IO will defer until the each chunk IO is reached.
+	  For contiguous storage, if we find the MPI-IO cannot
+	  support complicated MPI derived data type, we will
+	  set use_par_opt_io = FALSE.
+	*/
+    if(dset->shared->layout.type == H5D_CONTIGUOUS) {
+
+#ifndef H5_MPI_COMPLEX_DERIVED_DATATYPE_WORKS
+       if((H5S_SELECT_IS_REGULAR(file_space) != TRUE) ||
+	  (H5S_SELECT_IS_REGULAR(mem_space)  != TRUE))
+	   HGOTO_DONE(FALSE);
+#endif
+    }
     
-      /* Case 2:Check whether the number of chunks that covered the single hyperslab is the same.
-	 If not,no collective chunk IO. 
-	 KY, 2004/7/14
-      */
-	 
-        c1 = H5S_SELECT_IS_SINGLE(file_space);
-        c2 = H5S_SELECT_IS_SINGLE(mem_space);
-
-        if(c1==FAIL || c2 ==FAIL)
-            HGOTO_ERROR(H5E_DATASPACE, H5E_BADRANGE, FAIL, "invalid check for single selection blocks");
-        if(c1==FALSE || c2 ==FALSE)
-            HGOTO_DONE(FALSE);
-
-        /* Compute the number of chunks covered by the selection on this process */
-        tnum_chunkf = 1;
-        for (u=0; u<dim_rankf; u++) {
-            dim_chunks = (endf[u]/chunk_dim[u]-startf[u]/chunk_dim[u])+1;
-            tnum_chunkf = dim_chunks*tnum_chunkf;
-        }
-
-        /* Determine the minimum and maximum # of chunks for all processes */
-        if (MPI_SUCCESS != (mpi_code= MPI_Reduce(&tnum_chunkf,&max_chunkf,1,MPI_INT,MPI_MAX,0,comm)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_Reduce failed", mpi_code)
-        if (MPI_SUCCESS != (mpi_code= MPI_Reduce(&tnum_chunkf,&min_chunkf,1,MPI_INT,MPI_MIN,0,comm)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_Reduce failed", mpi_code)
-  
-        /* Let the rank==0 process determine if the same number of chunks will be operated on by all processes */
-        if(mpi_rank == 0)
-            num_chunks_same = (max_chunkf==min_chunkf);
-                    
-        /* Broadcast the flag indicating the number of chunks are the same */
-        if (MPI_SUCCESS != (mpi_code= MPI_Bcast(&num_chunks_same,1,MPI_INT,0,comm)))
-            HMPI_GOTO_ERROR(FAIL, "MPI_Bcast failed", mpi_code)
-
-        /* Can't handle case when number of chunks is different (yet) */
-        if(!num_chunks_same)
-            HGOTO_DONE(FALSE);
-    } /* end if */
+    if(dset->shared->layout.type == H5D_CHUNKED) 
+      if(dset->shared->dcpl_cache.pline.nused>0)
+	   HGOTO_DONE(FALSE); /* Perform the independent write operation */
+
+
 
 done:
     FUNC_LEAVE_NOAPI(ret_value);
+
 } /* H5D_mpio_opt_possible() */
 
 
@@ -370,281 +264,68 @@ done:
     FUNC_LEAVE_NOAPI(ret_value);
 } /* end H5D_mpio_spaces_xfer() */
 
-
-static herr_t
-H5D_mpio_spaces_span_xfer(H5D_io_info_t *io_info, 
-			  size_t elmt_size,
-			  const H5S_t *file_space, 
-			  const H5S_t *mem_space,
-			  void *_buf /*out*/, 
-			  hbool_t do_write )
-{
-    haddr_t	 addr;                              /* Address of dataset (or selection) within file */
-    size_t	 mpi_buf_count, mpi_file_count;     /* Number of "objects" to transfer */
-    hsize_t	 mpi_buf_offset, mpi_file_offset;   /* Offset within dataset where selection (ie. MPI type) begins */
-    MPI_Datatype mpi_buf_type, mpi_file_type;      /* MPI types for buffer (memory) and file */
-    hbool_t	 mbt_is_derived=0,                 /* Whether the buffer (memory) type is derived and needs to be free'd */
-		 mft_is_derived=0;                /* Whether the file type is derived and needs to be free'd */
-    hbool_t	 plist_is_setup=0;                /* Whether the dxpl has been customized */
-    uint8_t	*buf=(uint8_t *)_buf;             /* Alias for pointer arithmetic */
-    int          mpi_code;                        /* MPI return code */
-    herr_t	 ret_value = SUCCEED;             /* Return value */
-
-
-    FUNC_ENTER_NOAPI_NOINIT(H5D_mpio_spaces_span_xfer);
-
-    /* Check args */
-    assert (io_info);
-    assert (io_info->dset);
-    assert (file_space);
-    assert (mem_space);
-    assert (buf);
-    assert (IS_H5FD_MPIO(io_info->dset->ent.file));
-
-    /* Make certain we have the correct type of property list */
-    assert(TRUE==H5P_isa_class(io_info->dxpl_id,H5P_DATASET_XFER));
-
-
-    /* create the MPI buffer type */
-    if(H5S_SELECT_IS_REGULAR(mem_space)==TRUE){
-      if (H5S_mpio_space_type( mem_space, elmt_size,
-			       /* out: */
-			       &mpi_buf_type,
-			       &mpi_buf_count,
-			       &mpi_buf_offset,
-			       &mbt_is_derived )<0)
-    	HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buf type");
-    }
-    else {
-      if (H5S_mpio_space_span_type( mem_space, elmt_size,
-			       /* out: */
-			       &mpi_buf_type,
-			       &mpi_buf_count,
-			       &mpi_buf_offset,
-			       &mbt_is_derived )<0)
-    	HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI buf type");
-    }
-
-    /* create the MPI file type */
-
-   if(H5S_SELECT_IS_REGULAR(file_space)== TRUE){ 
-     if ( H5S_mpio_space_type( file_space, elmt_size,
-			       /* out: */
-			       &mpi_file_type,
-			       &mpi_file_count,
-			       &mpi_file_offset,
-			       &mft_is_derived )<0)
-    	HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI file type");
-   }
-   else {
-     if ( H5S_mpio_space_span_type( file_space, elmt_size,
-			       /* out: */
-			       &mpi_file_type,
-			       &mpi_file_count,
-			       &mpi_file_offset,
-			       &mft_is_derived )<0)
-    	HGOTO_ERROR(H5E_DATASPACE, H5E_BADTYPE, FAIL,"couldn't create MPI file type");
-   }
-
-   /* Get the base address of the contiguous dataset or the chunk */
-   if(io_info->dset->shared->layout.type == H5D_CONTIGUOUS)
-       addr = H5D_contig_get_addr(io_info->dset) + mpi_file_offset;
-   else {
-       haddr_t   chunk_addr; /* for collective chunk IO */
-       assert(io_info->dset->shared->layout.type == H5D_CHUNKED); 
-       chunk_addr=H5D_istore_get_addr(io_info,NULL);
-       addr = H5F_BASE_ADDR(io_info->dset->ent.file) + chunk_addr + mpi_file_offset;
-   }
-
-   /*
-    * Pass buf type, file type to the file driver. Request an MPI type
-    * transfer (instead of an elementary byteblock transfer).
-    */
-   if(H5FD_mpi_setup_collective(io_info->dxpl_id, mpi_buf_type, mpi_file_type)<0)
-        HGOTO_ERROR(H5E_PLIST, H5E_CANTSET, FAIL, "can't set MPI-I/O properties");
-   plist_is_setup=1;
-
-   /* Adjust the buffer pointer to the beginning of the selection */
-   buf+=mpi_buf_offset;
-
-   /* transfer the data */
-   if (do_write) {
-       if (H5F_block_write(io_info->dset->ent.file, H5FD_MEM_DRAW, addr, mpi_buf_count, io_info->dxpl_id, buf) <0)
-	   HGOTO_ERROR(H5E_IO, H5E_WRITEERROR, FAIL,"MPI write failed");
-    } 
-   else {
-    	if (H5F_block_read (io_info->dset->ent.file, H5FD_MEM_DRAW, addr, mpi_buf_count, io_info->dxpl_id, buf) <0)
-	    HGOTO_ERROR(H5E_IO, H5E_READERROR, FAIL,"MPI read failed");
-   }
-
-done:
-    /* Reset the dxpl settings */
-    if(plist_is_setup) {
-        if(H5FD_mpi_teardown_collective(io_info->dxpl_id)<0)
-    	    HDONE_ERROR(H5E_DATASPACE, H5E_CANTFREE, FAIL, "unable to reset dxpl values");
-    } /* end if */
-
-    /* free the MPI buf and file types */
-    if (mbt_is_derived) {
-	if (MPI_SUCCESS != (mpi_code= MPI_Type_free( &mpi_buf_type )))
-            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code);
-    }
-    if (mft_is_derived) {
-	if (MPI_SUCCESS != (mpi_code= MPI_Type_free( &mpi_file_type )))
-            HMPI_DONE_ERROR(FAIL, "MPI_Type_free failed", mpi_code);
-    }
-
-    FUNC_LEAVE_NOAPI(ret_value);
-} /* end H5D_mpio_spaces_span_xfer() */
-
 
 /*-------------------------------------------------------------------------
- * Function:	H5D_mpio_spaces_read
+ * Function:	H5D_mpio_select_read
  *
  * Purpose:	MPI-IO function to read directly from app buffer to file.
  *
  * Return:	non-negative on success, negative on failure.
  *
- * Programmer:	rky 980813
+ * Programmer:	
  *
  * Modifications:
  *
- * rky 980918
- * Added must_convert parameter to let caller know we can't optimize the xfer.
- *
- *      QAK - 2002/04/02
- *      Removed the must_convert parameter and move preconditions to
- *      H5S_mpio_opt_possible() routine
- *
  *-------------------------------------------------------------------------
  */
 herr_t
-H5D_mpio_spaces_read(H5D_io_info_t *io_info,
+H5D_mpio_select_read(H5D_io_info_t *io_info,
     size_t UNUSED nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
     void *buf/*out*/)
 {
     herr_t ret_value;
 
-    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_spaces_read);
+    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_select_read);
 
-    ret_value = H5D_mpio_spaces_xfer(io_info, elmt_size, file_space,
+           
+      ret_value = H5D_mpio_spaces_xfer(io_info, elmt_size, file_space,
         mem_space, buf, 0/*read*/);
 
     FUNC_LEAVE_NOAPI(ret_value);
-} /* end H5D_mpio_spaces_read() */
+} /* end H5D_mpio_select_read() */
 
 
 /*-------------------------------------------------------------------------
- * Function:	H5D_mpio_spaces_write
+ * Function:	H5D_mpio_select_write
  *
  * Purpose:	MPI-IO function to write directly from app buffer to file.
  *
  * Return:	non-negative on success, negative on failure.
  *
- * Programmer:	rky 980813
+ * Programmer:	
  *
  * Modifications:
  *
- * rky 980918
- * Added must_convert parameter to let caller know we can't optimize the xfer.
- *
- *      QAK - 2002/04/02
- *      Removed the must_convert parameter and move preconditions to
- *      H5S_mpio_opt_possible() routine
  *
  *-------------------------------------------------------------------------
  */
 herr_t
-H5D_mpio_spaces_write(H5D_io_info_t *io_info,
+H5D_mpio_select_write(H5D_io_info_t *io_info,
     size_t UNUSED nelmts, size_t elmt_size,
     const H5S_t *file_space, const H5S_t *mem_space,
     const void *buf)
 {
     herr_t ret_value;
 
-    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_spaces_write);
+    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_select_write);
 
     /*OKAY: CAST DISCARDS CONST QUALIFIER*/
-    ret_value = H5D_mpio_spaces_xfer(io_info, elmt_size, file_space,
-        mem_space, (void*)buf, 1/*write*/);
+       ret_value = H5D_mpio_spaces_xfer(io_info, elmt_size, file_space,
+                    mem_space, (void*)buf, 1/*write*/);
 
     FUNC_LEAVE_NOAPI(ret_value);
 } /* end H5D_mpio_spaces_write() */
 
 
-
-/*-------------------------------------------------------------------------
- * Function:	H5D_mpio_spaces_span_read
- *
- * Purpose:	MPI-IO function to read directly from app buffer to file for
-                span-tree
- *
- * Return:	non-negative on success, negative on failure.
- *
- * Programmer:	KY
- *
- * Modifications:
- *
- *
- *-------------------------------------------------------------------------
- */
-herr_t
-H5D_mpio_spaces_span_read(H5D_io_info_t *io_info,
-			  size_t UNUSED nelmts, 
-			  size_t elmt_size,
-			  const H5S_t *file_space, 
-			  const H5S_t *mem_space,
-			  void *buf/*out*/)
-{
-    herr_t ret_value;
-
-    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_spaces_span_read);
-
-    ret_value = H5D_mpio_spaces_span_xfer(io_info, elmt_size, file_space,
-        mem_space, buf, 0/*read*/);
-
-    FUNC_LEAVE_NOAPI(ret_value);
-} /* end H5D_mpio_spaces_read() */
-
-
-/*-------------------------------------------------------------------------
- * Function:	H5D_mpio_spaces_span_write
- *
- * Purpose:	MPI-IO function to write directly from app buffer to file.
- *
- * Return:	non-negative on success, negative on failure.
- *
- * Programmer:	KY
-                 
- *
- * Modifications:
- *
- * rky 980918
- * Added must_convert parameter to let caller know we can't optimize the xfer.
- *
- *      QAK - 2002/04/02
- *      Removed the must_convert parameter and move preconditions to
- *      H5S_mpio_opt_possible() routine
- *
- *-------------------------------------------------------------------------
- */
-herr_t
-H5D_mpio_spaces_span_write(H5D_io_info_t *io_info,
-			   size_t UNUSED nelmts, 
-			   size_t elmt_size,
-			   const H5S_t *file_space, 
-			   const H5S_t *mem_space,
-			   const void *buf)
-{
-    herr_t ret_value;
-
-    FUNC_ENTER_NOAPI_NOFUNC(H5D_mpio_spaces_span_write);
-
-    /*OKAY: CAST DISCARDS CONST QUALIFIER*/
-    ret_value = H5D_mpio_spaces_span_xfer(io_info, elmt_size, file_space,
-        mem_space, (void*)buf, 1/*write*/);
-
-    FUNC_LEAVE_NOAPI(ret_value);
-} /* end H5D_mpio_spaces_span_write() */
 #endif  /* H5_HAVE_PARALLEL */